Research On Large-Scale Isolated Word Recognition Algorithm

At present, the performance of the isolated word recognition system with a small and medium-sized list has almost fulfilled the requirement of people. However, with the growing of the size of the dictionary, the performance of the isolated word recognition system goes down sharply because of the inflation of the search space. Therefore, the large-scale isolated-word recognition system needs to be studied further. In this paper, we will have a discussion about the problems we faced and the achievements we got when we did research on the large-list isolated-word recognition system.1. In this paper, a framework based on a two-level recognition network was designed to alleviate the problem of the inflation of the size of the dictionaries. In addition, an algorithm based on the prefix tree was proposed to calculate the edit distance. Compared with the traditional algorithm used to calculate edit distance, this algorithm is more effective and less sensitive to the size of the wordlist.2. A pruning method based on the number of states possessed by the hypotheses was proposed to improve the performance of the search process. The pruning method based on the number of states possessed by the hypotheses could improve the performance of the ASR obviously.3. At the same time, in the first level, a pruning algorithm based on the back off weights of the language model was proposed. And the speed of the first level was improved obviously.4. In addition, the recognition network based on WFST is studied and built, a comparison between the network based on WFST and the network based on the prefix tree was also made.